Mender auto-fix

Mender is Guard's AI remediation layer. When a scenario fails — on a PR, in production, or during shadow-prod — Mender triages the failure, classifies root cause, and optionally opens a fix PR against your application repository. Mender targets your product code, not just the test file.

Human review is always required. Mender never merges its own PRs.

Product page: guard.molar.it#mender

The Mender loop

failed run
    → triage (classify + confidence score)
    → branch by classification
        flake     → retry 3× with backoff
        env       → alert Molar ops
        scenario_bug → self-heal PR against .molar/scenarios/
        product_bug  → fix-PR agent → sandbox validation → GitHub PR
    → human merge (or reject → RLHF signal)
    → Guard re-runs on fix PR → green check

Triage stage

Every failed run enters triage before any code is written. Misclassifying a flake as product_bug annoys engineers; misclassifying a product bug as flake is dangerous.

Classifier inputs

Classifier output

{
  "classification": "product_bug",
  "confidence": 0.87,
  "reasoning": "POST /api/subscriptions/upgrade reads req.body.priceId but field renamed in PR #4521",
  "suspected_files": ["app/api/billing/route.ts"],
  "suspected_pr": "https://github.com/org/repo/pull/4521"
}

Actions by classification

Confidence scores

Confidence is a 0..1 score stored on every mender_attempts row. Use it in the dashboard Mender inbox to prioritize review:

In suggestive mode, low-confidence fixes appear as comments or dashboard suggestions only — not auto-PRs.

Fix-PR stage (product bugs)

Step 1 — Context gathering

The agent assembles:

• Failed scenario file (verbatim)
• Assertion + stack + screenshots + DOM diff vs last green run
• HAR of failed run
• Suspected file contents (via GitHub API)
• CONTRIBUTING.md, .cursorrules, AGENTS.md if present
• Repository map (tree-sitter symbol index)

Step 2 — Patch generation

• Agent loop: up to 20 LLM calls per attempt (hard cap against token snowball)
• Tools: read_file, grep, apply_diff, run_scenario_in_sandbox, get_repo_map
• Patch format: unified diff (aider-style applier)
• Validation: Run failing scenario in Docker sandbox with Clones — not prod, not staging. Max 3 outer iterations.

Step 3 — PR creation

PR body includes:

• Failing scenario link + region + timestamp
• Suspected cause (LLM reasoning)
• Inline diff
• Local verify command: molar run <slug> --against=this-branch
• Sandbox validation result
• Token usage + model metadata
• Reject this fix link (RLHF)

Step 4 — Loop closure

Push branch → PR webhook → Guard runs affected scenarios → green check → human merges.

Mender modes

Configure per repo (molar-guard.config.ts) or per scheduled check:

// molar-guard.config.ts
export default {
  mender: {
    mode: "suggestive",
    forbiddenPaths: [
      "app/api/auth/**",
      "**/payment*",
      "**/security*",
      "migrations/**",
    ],
  },
};

Suggestive mode actions

• PR-time failure: Patch preview on the failing PR comment with Approve / Reject / Edit
• Production failure: Dashboard incident panel → Apply fix → converts suggestion to real GitHub PR
• API: POST /v1/mender/:id/apply

Safety & governance

Use BYOK for Mender inference — customer keys bill upstream directly; Molar-hosted caps do not apply.

Reject and accept signals (learning loop)

Reject

Close PR without merging, click Reject this fix in PR body, or call:

curl -X POST https://api.molar.it/v1/mender/$ATTEMPT_ID/reject \
  -H "Authorization: Bearer $MOLAR_API_KEY" \
  -d '{"reason": "Wrong root cause — auth middleware issue, not billing"}'

Reasons feed RLHF for prompt improvement. Optional taxonomy: wrong file, wrong fix, intentional change, flake misclassified.

Accept

When a Mender PR merges, Guard stores the final diff (including human edits). The delta between Mender's original patch and merged code is high-value training signal.

Mender dashboard

At app.molar.it/dashboard/guard → Mender (or Fixbots in combined bundle):

Inbox list:

• Filter: pending review, classification, outcome, repo
• Columns: classification, confidence, scenario, PR URL, token cost, outcome

Attempt detail:

• Triage rationale + evidence attachments
• Diff preview (GET /v1/mender/:id/preview)
• Reject with reason
• Apply (suggestive → real PR)
• Budget + circuit breaker status

MCP tools for agents

Standalone Guard MCP (stdio transport) is available for self-hosted deployments:

The unified 22-tool Molar MCP connector (mcp.molar.cloud) proxies additional tools via the Molar API. See Webhooks & API.

Model routing

Mender uses separate models per capability (configurable in Settings → AI):

Default is Molar-hosted; BYOK supported per BYOK docs.

What Mender is not

• Not a merge bot — humans approve every fix
• Not silent test healing — scenario changes always arrive as PRs
• Not a replacement for code review — patches can be wrong; confidence scores guide priority
• Not allowed on forbidden paths — payment/auth/security/migrations alert only

Next

• GitHub App setup — Mender triggers from PR failures
• Production monitoring — Mender on prod incidents
• Security — Mender governance and prompt-injection controls
• Troubleshooting — circuit breaker and budget issues